Re: Clarifying platform_device_unregister

[Richard Purdie <rpurdie@rpsys.net>]

On Tue, 2008-04-01 at 10:54 -0400, Dmitry Torokhov wrote:
> On Tue, Apr 01, 2008 at 12:47:54AM -0700, Jaya Kumar wrote:
> > That's interesting. I noticed though that a lot of platform device
> > code assigns a statically allocated structure to platform_data. For
> > example:
> > 
> > arch/arm/mach-pxa/corgi_pm.c
> > static struct sharpsl_charger_machinfo corgi_pm_machinfo = {
> > ...
> > }
> >         corgipm_device->dev.platform_data = &corgi_pm_machinfo;
> > 
> > same with spitz_pm.c.
> > 
> > egrep "platform_data.*=.*\&" *.c shows quite a lot of users doing
> > that. I guess most of these below are probably okay since these
> > drivers can't be rmmoded.
> > 
> 
> Hmm, are you sure they can't be removed? Why do they all have
> module_exit methods?
> 
> Even if they can't be unloaded the whole thing will blow to pieces
> if registration fails. Consider this:
> 
> static int __devinit spitzpm_init(void)
> {
>         int ret;
> 
>         spitzpm_device = platform_device_alloc("sharpsl-pm", -1);
>         if (!spitzpm_device)
>                 return -ENOMEM;
> 
>         spitzpm_device->dev.platform_data = &spitz_pm_machinfo;
>         ret = platform_device_add(spitzpm_device);
> 
>         if (ret)
>                 platform_device_put(spitzpm_device);
> 		^^^^^^^^^^^
> This will try to kfree(spitzpm_device->dev.platform_data) and it gonna
> blow. We need to do spitzpm_device->dev.platform_data = NULL before doing
> put.
> 
> Also  spitzpm_init() shoudl be marked __init, not __devinit and
> spitzpm_exit() should be __exit() if it is event needed at all.
> 
> Richard, I think you work with spitz and corgi, any comments?

Looking at this I agree there are some problems there.

In the real world the spitz/corgi devices are pretty useless without
power management and therefore these code paths aren't well used. I
think this has happened since the platform data was a later addition to
the code and the internal use of kfree on platform_data wasn't realised.

I'll make sure some patches are submitted to address these issues.

Cheers,

Richard